Magic TeamsHow to Plan Agency Capacity With AI
To plan agency capacity with AI, put four things in one weekly view: the work you’ve committed to, the hours each person can realistically give, the dependencies that could move that work, and a short list of exceptions a human has to decide on. AI does the gathering, the estimate correction and the flagging. A person decides who moves, what slips and what the client hears.
Here’s the situation this fixes. It’s Monday. Your project tool says the design team has room, your resourcing spreadsheet says they’re at 90%, and your senior designer says she’s drowning.
All three are partly right. None of them sees the whole picture.
Most agencies don’t lack data here. They lack one place where the data meets, and one person who acts on it.
Why is agency capacity planning so hard to get right?
The inputs live in different systems and go stale at different speeds. Scopes sit in proposals. Estimates sit in the project tool. Time off sits in a calendar. Client delays sit in someone’s inbox.
Teamwork.com’s agency operations benchmarking report (the 2023 edition, a survey of more than 500 agency leaders) found that 52% of agencies struggle to push billable utilization above 50%. Only 18% always hit their billable vs. non-billable time targets. Just 33% said their people submit time-tracking data on time about 90% of the time.
Teamwork sells agency software, so treat those as vendor-survey figures. They still match what most agency owners describe: the time data is patchy, so the plan built on it is too.
Why can’t you just trust the estimates?
Because people underestimate their own work, and they do it consistently.
In a classic study, Buehler, Griffin and Ross (1994) asked honors students when they’d finish their theses. Their best estimate averaged 33.9 days. They actually took 55.5. Fewer than a third (29.7%) finished by their best guess, and fewer than half (48.7%) finished even by the date they gave assuming everything went as badly as possible.
Here’s the useful part. The same study found the estimates were still strongly correlated with actual time (r = .77). People who said a thesis would take longer did take longer.
That’s students, not agencies. But the planning fallacy shows up wherever people estimate their own work, and the finding points at a fix: estimates are biased but informative. So don’t throw them out. Correct them with your own history.
That correction is the single most useful thing AI does in capacity planning.
What should an AI capacity plan include?
We use a structure we call the Capacity Board. It’s the shape we design when we build a weekly planning workflow into a client’s AI operating system. It came from one design question: what does a founder or ops lead need to see to make next week’s staffing call in ten minutes?
The answer kept coming back to four inputs, read in this order:
- Committed work. Everything sold or promised, with an hour estimate per person. AI corrects each estimate using your actuals on similar past work, by task type. If landing pages really take you 1.25x the estimate, a “20-hour” landing page gets planned at 25.
- Realistic availability. Contract hours minus time off, recurring meetings, internal admin and standing non-billable duties. This is the number teams overstate most.
- Dependencies. Client copy, approvals, assets and internal handoffs that must arrive before work can start. Each gets a due date and an owner.
- Exception review. The system doesn’t hand you the whole board. It hands you the handful of rows that break a rule you set.
That last step is where most of the value sits. If you have to read the board line by line, you’ve built a nicer spreadsheet.
What does AI do, and what stays with a person?
The split matters more than the tool. Here’s where we draw the line:
| Task | AI’s role | Person’s role |
|---|---|---|
| Pull scopes, estimates and time off into one view | Does it | Checks the sources are right |
| Correct estimates using past actuals | Does it and shows the ratio used | Can override per project |
| Spot overloaded or idle people | Flags them | Decides what to do |
| Track client dependencies | Drafts the reminder | Sends or edits it |
| Reassign work between people | Suggests options | Makes the call |
| Change a client deadline | Never on its own | Decides and talks to the client |
| Judge someone’s performance | Not from this data | Not from this board alone |
The last row is a real boundary. Capacity data shows load. It doesn’t show effort or quality.
If you plan to use time or activity data in performance reviews, employment and privacy rules vary by country and state. Talk to an HR or employment-law advisor first.
What does this look like in a real week?
This worked example is hypothetical. The agency, people and numbers are made up to show the math. They aren’t client results.
A 12-person agency has a three-person design team. Next week they’ve estimated 70 hours of design work across four clients. Their own records show design tasks take about 1.25x the estimate, so the board plans for 87.5 hours.
Availability:
- Ana: 40 contract hours, minus 8 for meetings and admin = 32 available.
- Ben: 40 hours, minus 16 for two days off, minus 4 admin = 20 available.
- Cara: 40 hours, minus 8 for meetings and admin = 32 available.
That’s 84 hours available against 87.5 committed. At team level, 104%. Tight, nothing alarming.
Per person, it’s a different story. The chart compares each designer’s raw estimate, the corrected estimate and their real availability:
On raw estimates, everyone looks fine. Ana sits at 90% and Cara at 94%. After the correction, Ana is at 113% and Cara at 117%, while Ben has six free hours.
A spreadsheet of raw estimates would have shown a green week. A team total would have hidden who was over.
Then the dependency check fires. Fifteen of Cara’s corrected hours are the Client A homepage, which needs final copy from the client by Wednesday. The copy hasn’t arrived, and the last two copy deliveries from this client were late.
So Monday’s exception list has four lines:
- Cara is at 117%, and 15 hours depend on late-prone client copy.
- Ana is at 113%.
- Ben has 6 open hours.
- Client A copy is due Wednesday, with a draft reminder ready.
The ops lead spends ten minutes on it. They move a 5-hour Client C task from Ana to Ben, which puts Ana at about 97% and Ben at 95%. They send the copy reminder today instead of Wednesday.
Cara is the harder call. The ops lead tells Client A that if copy lands after Wednesday, the homepage moves to the following week. That would drop Cara to 22.5 hours, about 70%. If the copy arrives on time, they approve a few freelancer hours on a smaller task so Cara isn’t carrying 117% all week.
Notice what the AI didn’t do. It didn’t reassign anyone, and it didn’t tell the client anything. It made the decision visible early enough to be cheap.
How do you set the exception rules?
Start with three or four rules you’d actually act on. Too many flags and people stop reading them.
Treat those thresholds as starting points you’ll tune, not industry benchmarks. Some agencies plan billable roles to a lower ceiling on purpose, to leave room for reactive client work. Set yours from how your team actually runs.
When does the board tell you to hire?
One overloaded week is a scheduling problem. Several weeks in a row where the corrected load sits above realistic availability, across the team and not just one person, is a capacity problem.
That’s when the decision moves from “who takes this task” to “freelancer, hire, or say no to some work.” The board doesn’t make that call. It gives you a few weeks of honest numbers to make it with.
Before you hire, check whether the overload traces back to one person approving everything. If it does, more people won’t fix it. We wrote about that in why hiring more people won’t fix your bottleneck.
What do you need before you start?
You don’t need perfect data. You need enough to make the correction honest.
- Active projects with an owner and an hour estimate per person
- At least a few weeks of actual hours on similar task types
- Time off and recurring meetings in a calendar the system can read
- Client dependencies written down with a due date
- Three or four exception rules your ops lead agrees with
- One named person who reviews the exception list every week
If your time data is patchy, say so on the board. A 1.25x correction built on four weeks of half-filled timesheets deserves less trust than one built on a year of clean data. Show the confidence next to the number.
When we design planning workflows at Magic Teams, we design the review step before the dashboard. Our reasoning: if nobody owns the Monday exception list, the board turns into another report people skim. We’d rather ship a plain list with a named reviewer than a polished chart nobody acts on.
How do you know it’s working?
Pick two or three measures before you start and check them monthly:
- How close planned hours land to actual hours, by person.
- How many deadlines moved because of a surprise vs. a decision made early.
- How many person-weeks ran above 100%.
Don’t credit the board for every good month. Seasonality and client mix move these numbers too. For a cleaner way to separate real gains from noise, see how to measure ROI on AI automation.
FAQ
Do I need detailed time tracking for this to work?
It helps, but you can start without it. Use project-level actuals or a weekly “how long did this really take” check-in. The correction gets better as the data does.
What utilization target should I plan to?
There isn’t one right number. It depends on role, pricing model and how much reactive work your clients send. Plan to a ceiling below 100% of realistic hours, then adjust after a few months of comparing plan to actual.
Can AI forecast capacity months ahead?
It can project from signed work and weighted pipeline, but treat anything past a few weeks as a range. Deals slip and scopes change. The weekly board is where decisions get made.
Which tools do I need?
Your existing project tool, a shared calendar and something that can read both and run the rules. The structure and the review habit matter more than the product.
Will this help if I’m the bottleneck?
Partly. The board makes the overload visible. If every exception still routes to you, the next step is handing someone else authority over the review. See how to stop being the bottleneck in your business.
Who wrote this?
This guide is by Satya Phanindra Reddy, founder of Magic Teams AI. We install local-first, human-in-the-loop AI operating systems for founder-led agencies and small professional practices, usually in a one-week intensive.
The Capacity Board comes from that design work: deciding which parts of weekly planning need a machine’s memory and math, and which need a person’s judgment.
If your team’s capacity lives in three tools and one person’s head, and you want a planning workflow your team can review each week, that’s a good topic for a fit call with Magic Teams.